I am new to the Movidius Stick platform, before posting the question, I have gone through the similar discussions in this Forum, but I am not able to solve the following problem that is appearing:
After giving the following command:
mvNCCompile ckpt/ckpt.meta -in=input_1 -on=bottleneck/Elu
[Error 4] Toolkit Error: Stage Type Not Supported: ExpandDims
mvNCCompile ckpt/ckpt.meta -in=input_1 -on=bottleneck/Elu --old-parser
[Error 5] Toolkit Error: Stage Details Not Supported: VarHandleOp
I have already generated my meta file using Tensorflow Keras, but not able to use the mmvNCCompile with this meta file.
I am attaching the snapshot of the operations, where I am stuck.
Would you be able to help me on this again?
Could you send a link to your model? It would greatly help us solve this issue.
The "Stage Details Not Supported:..." error may indicate that there were unsupported layers or operations involved in the conversion of the custom trained model to Tensorflow and/or in the conversion to the NCS graph.
Oh ! Seems I missed the file.
I am attaching the Model.zip containing all the info.
I have an update, I tried using OpenVINO yesterday, as I came to know that OpenVINO support more layers and have better dev support.
I have stuck using OpenVINO to, the logs are here:
[ ERROR ] Run Model Optimizer with --log_level=DEBUG for more information. [ ERROR ] Exception occurred during running replacer "REPLACEMENT_ID" (<class 'extensions.middle.PartialInfer.PartialInfer'>): Stopped shape/value propagation at "AssignVariableOp_7" node. For more information please refer to Model Optimizer FAQ (<INSTALL_DIR>/deployment_tools/documentation/docs/MO_FAQ.html), question #38.
Dear Aroop, thanks for the debugging,
I am sorry that you couldn't reproduce the error.
Actually, I am not using the *.pb file, I am using the meta file for generating the graph.
As it was suggested at the following tutorial:
Detailed procedure below:
1.) We have obtained the trained model named "core3_ac_model.h5" using the Keras from Ternsorflow version 1.13.1
2.) We converted the Keras model to Tensoflow model and obtained following four files:
Checkpoint file: checkpoint
Checkpoint data file: ckpt.data-00000-of-00001
Index file: ckpt.index
Meta file: ckpt.meta
Please note: I had shared only the meta file before, as this was the only needed as per the following tutorial of OpenVINO:
3.) Now, using OpenVINO version 2019-R1.1, I am trying to obtain the graph file for Movidius
I am using the following command:
python3 mo_tf.py --input_meta_graph ~/ckpt/ckpt.meta --log_level=DEBUG
I am getting the following error on console: ( I am attaching the whole log as attachment)
I noticed that you also posted on the Computer Vision Forum here: https://software.intel.com/en-us/forums/computer-vision/topic/806158
The error seems to be caused by a known bug in the current release (OpenVino 2019 R1.1).
Thanks for the continuous support.
With the help, I could proceed to generate the Movidius Graph file, But this file is not compatible now when I use the Movidius USB3.0 NCS.
I would like to know, without the NCS USB3.0 Movidius device, can I test this graph file on a stand-alone PC?
Like in a simulation mode?
Assuming you are using OpenVINO and have converted your model to an IR format (.xml and .bin) you can test the model with one of our sample applications and specifying the target device with -d CPU .
During the model conversion, the model will be converted to FP32 by default, so this should work on your CPU. If you need to inference the model with a Neural Compute Stick, you will need to specify FP16 as the datatype.
Hope this answers your question.